AMI · The intelligence layer
Ask it, or put it inside the workflow.
An assistant your team can talk to, and eight steps any workflow can call. It reads the request, searches the knowledge you have published, and stops exactly where your rules tell it to.
- Grounded in
- Knowledge you have published
- Shown live
- Only the stages it actually ran
- Scope
- One workspace — the caller's own
The same intelligence,
reached two ways.
One is a conversation an employee starts. The other is a step a workflow runs at three in the morning. They read the same knowledge, under the same rules.
Way one
Ask it.
A workspace of its own, with every conversation on its own URL. It answers from what you have published, and lists what it used underneath — so an employee can check it rather than trust it.
Summarise #4188 and tell me what we promised the customer.
Marek reported an SSO redirect loop on Tuesday. Tier 2 confirmed the identity provider was returning a stale assertion and told him a fix would ship this week.
From your knowledge base
- SSO and identity providersArticle
- What we commit to on a Tier 2 escalationFAQ
Way two
Or put it inside the workflow.
Eight steps the workflow builder offers next to every other action. They take configuration like any other step, and the ones that can fail say so with a branch rather than guessing.
Let AMI check the knowledge base
Look for an existing answer before anyone spends time on the question.
- Only count answers above
- 78%
- Send the article to the customer
What happens between the question and the answer.
Seven stages, streamed as the work starts. Follow them, or let it play.
In
A question from an employee, or a step in a running workflow.
Out
An answer with its sources, a draft waiting for review, or the branch the workflow takes next.
Stage 01
Screened, then planned
A probe for a key, the system prompt or the provider underneath is answered here, from a constant — before a plan exists and before a single row of company data is read. What survives becomes a plan: in scope or not, knowledge or not, which reports, what date range.
- Runs before
- Any retrieval
- A plan is
- A choice among named reports
- Out of scope
- Answered, not attempted
A stage is announced immediately before the work it names begins, and nothing is announced for work that is skipped — so a question needing no statistics never shows “Analysing your support data”. None of the seven exposes model reasoning. They are pipeline stages.
An assistant is only as good as what it will not say.
Everything AMI can read, and the questions it answers from a constant rather than from a model.
Reaches
Your workspace
The ticket or chat in front of the employee, the customer on it, their team — scoped to the signed-in company and user id.
Your published knowledge
Articles and FAQs from your own knowledge base. Five per answer, 1,600 characters of each.
Your resolved conversations
Threads your team has already closed, read as precedent for tone and outcome.
Five named reports
A fixed set, chosen from an enum. Seven days by default, 180 at most, and never a query it composed itself.
Refuses, without asking a model
Another company on Epkora
Including whether one is a customer. The answer is the same sentence whoever you name — asking it a thousand times reveals a thousand copies of one sentence.
Its own keys and configuration
Credentials, environment, internal instructions and the provider underneath are answered before a plan exists. The request never becomes a model call.
Epkora as a business
Revenue, headcount and customer numbers are in no prompt it receives, so any answer would be invented — and an invented figure about the vendor, said inside the vendor’s product, reads as official.
Instructions hidden in your content
Retrieved articles and transcripts are wrapped and declared untrusted. Anything giving orders inside them is reported, not obeyed.
Eight steps, sitting beside every other action.
No separate automation product, no connector. AMI steps configure like any other step and branch like any other check.
Let AMI answer
Replies from your knowledge base and hands over the moment it is out of its depth.
AMI handled itNeeds a person
Let AMI check the knowledge base
Look for an existing answer before anyone spends time on the question.
Found an answerNothing useful
Let AMI sort the ticket
Reads the request and picks the right category, from the list you allow.
Let AMI work out what they want
Detects the intent behind the message — a refund, a bug report, a question.
Let AMI suggest a priority
Weighs the wording, the customer and the impact, under a ceiling you set.
Let AMI summarise the conversation
Drops a recap on the ticket so whoever picks it up next is up to speed.
Let AMI draft a reply
Prepares a response for the agent to review. Nothing is sent without approval.
Let AMI read the mood
Spots frustrated customers early, so you can step in before they churn.
Customer is unhappyAll good
A workflow that uses four of them
The failure has a path too.
Every AMI step that can come up empty says so with a branch instead of guessing. Below a confidence floor you choose, the run takes the human path — and the manager who authored the flow can read back every step of it afterwards as a timeline.
Trigger
Ticket created
- 01
Let AMI sort the ticket
Save the category on the ticket · Billing, Technical, Account
- 02
Let AMI read the mood
Treat as unhappy below 2/5
Customer is unhappyEscalate to a managerAll goodContinues below - 03
Let AMI check the knowledge base
Only count answers above 78% confidence
Found an answerSend the article to the customerNothing usefulContinues below - 04
Let AMI draft a reply
Professional · include knowledge base links
A draft, waiting for an agent
with the articles it used attached
You set the edges. It stays inside them.
Every one of these is a field in the product, not a policy document.
Confidence floor
Only count answers above a percentage you choose
Below it, the run takes the human path instead of the article.
Hand over after
A number of replies you choose
Or the moment it is out of its depth, whichever comes first.
Introduce itself
Say that the customer is talking to an assistant
On or off, per workflow.
Tone
Friendly · Professional · Short and to the point
Categories
Only the ones you allow
AMI picks from your list. It cannot invent a category that is not on it.
Priority ceiling
Never go higher than
And you decide separately whether the suggestion is applied at all, or left as a suggestion.
Unhappy below
A threshold out of five
Under it, the run takes the escalation branch before the customer asks for a manager.
Draft, or reply
Two different steps
A draft waits for an agent. Replying to a customer directly is a step you add on purpose, with a turn cap and a handover.
Four numbers you can go and count.
These are counts of things that exist in the product — items in a step picker, stages in a pipeline, entries in a fixed list. Open a trial and count them yourself. None of them is a statistic about anybody’s support desk, and this page does not have any.
8
AMI steps in the workflow builder
7
named stages, streamed while it works
5
knowledge results behind one answer
5
reports it may run, chosen from a fixed list
And three rules, which are not numbers
A drafted reply waits.
“Let AMI draft a reply” prepares a response for an agent and stops. Sending one to a customer is a different step, turned on deliberately, with a turn cap and a handover.
Answers cite what they used.
When a turn read your knowledge base, the articles and FAQs are listed underneath with links. When it read none — a statistics question, say — it lists none, rather than dressing the answer in sources it did not use.
Out of its depth is a branch.
Not a guess, and not silence. Below the floor you set, the workflow follows the path you drew for exactly that case.